Source-Trace

Problem

Ask an AI assistant a question and you get confident prose with a few citation chips scattered through it. Some sentences are backed by a source. Others only sound as if they are. Telling them apart means opening every link and reading it, which almost nobody does.

Solution

Source-Trace reads the answer directly on Perplexity or ChatGPT and marks every claim in place. Green means a citation sits on it, so go and check it says what is claimed. Orange means a source is cited but could not be verified, or it is the only one. Red means the sentence stands on nothing you can check.

What it never does is tell you something is true or false. It shows you where the sourcing is thin and helps you close the gap: search for the claim, find a second source, or run a deep trace that looks for independent sources on the web and sums up what each of them says.

Approach

One rule guided every feature: coach, don't judge. Each label describes the sourcing you can see, never the truth of the claim, and that holds for the deep trace too.

The second rule was never to make the reader wait. Marks appear instantly from checks that run in the browser, then get richer as the server answers. If the server is slow or down, the extension still works, just with less detail. A privacy mode goes further and keeps everything on the device.

Under the hood, the extension draws its interface in an isolated layer so it never disturbs the page it sits on, and a small Python service extracts claims and checks citations. Both sides share one written contract for the report, so they cannot quietly drift apart.

Trade-offs & bottlenecks

Every AI site needs its own reader to pick the answer out of the page, and a redesign on their side can break it overnight.

The deep trace relies on a language model with web search, which costs money on every call. Privacy mode switches it off, and that is the price of keeping everything local.